diff --git a/browser_use/agent/service.py b/browser_use/agent/service.py
index 8fd81ccdd..783694533 100644
--- a/browser_use/agent/service.py
+++ b/browser_use/agent/service.py
@@ -185,6 +185,8 @@ class Agent(Generic[Context, AgentStructuredOutput]):
 		include_tool_call_examples: bool = False,
 		**kwargs,
 	):
+		if not isinstance(llm, BaseChatModel):
+			raise ValueError('invalid llm, must be from browser_use.llm')
 		# Check for deprecated planner parameters
 		planner_params = [planner_llm, use_vision_for_planner, is_planner_reasoning, extend_planner_system_message]
 		if any(param is not None and param is not False for param in planner_params) or planner_interval != 1:
diff --git a/browser_use/llm/base.py b/browser_use/llm/base.py
index 6bc4ae563..0e623d780 100644
--- a/browser_use/llm/base.py
+++ b/browser_use/llm/base.py
@@ -4,7 +4,7 @@ We have switched all of our code from langchain to openai.types.chat.chat_comple
 For easier transition we have
 """
 
-from typing import Any, Protocol, TypeVar, overload
+from typing import Any, Protocol, TypeVar, overload, runtime_checkable
 
 from pydantic import BaseModel
 
@@ -14,6 +14,7 @@ from browser_use.llm.views import ChatInvokeCompletion
 T = TypeVar('T', bound=BaseModel)
 
 
+@runtime_checkable
 class BaseChatModel(Protocol):
 	_verified_api_keys: bool = False